The impact of child labor on schooling outcomes in Nicaragua

Economics of Education Review 30 (2011) 1527–1539

Contents lists available at SciVerse ScienceDirect

Economics of Education Review journal homepage: www.elsevier.com/locate/econedurev

The impact of child labor on schooling outcomes in Nicaragua Mariela Buonomo Zabaleta ∗ Education For All Global Monitoring Report, UNESCO, France

a r t i c l e

i n f o

Article history: Received 19 November 2008 Received in revised form 18 July 2011 Accepted 11 August 2011 JEL classification: J22 J24 O15 Keywords: Child labor Educational economics Human capital

a b s t r a c t Child labor is considered a key obstacle to reaching the international commitments of Education For All. However, the empirical evidence on the effects of child labor on educational attainments is mostly limited to static measurements. This paper assesses the consequences of child labor on schooling outcomes over time by employing a three-year longitudinal household data set from Nicaragua. The potential endogeneity of past child labor and school outcomes is addressed through instrumental variables. The time a child dedicates to work is found to have harmful consequences on subsequent educational achievements, even after controlling for previous human capital accumulation and other factors. In particular, working over three hours a day is associated with school failure in the medium term. A distinction by type of work shows that time spent in market production has larger negative effects on school outcomes than time spent performing household chores. © 2011 Elsevier Ltd. All rights reserved.

1. Introduction and literature review This paper aims to identify the effect of children’s work on their educational attainments in the medium term by employing a panel data set from Nicaragua. The substantial benefits derived from education are well documented in their individual and societal dimensions, but less is known about the causal relationship between child labor and schooling. Underlying many recommendations of international bodies is the assumption that children’s work hinders educational achievement, and, in fact, there are several means through which children’s involvement in labor activities might interfere with their human capital formation. However, the economic theory does not predict unambiguously that child labor displaces schooling and most of the empirical evidence uses a static approach.

∗ Permanent address: Concepción del Uruguay 1615, Ap. 101, 11400 Montevideo, Uruguay. Tel.: +598 26191619. E-mail address: [email protected] 0272-7757/$ – see front matter © 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.econedurev.2011.08.008

Understanding how children’s work affects their educational attainments is especially important in the context of the educational goals set by the international community for 2015. Nicaragua has made significant progress in terms of primary education coverage. As shown in Table 1, net enrolment rates in this level increased 20 percentage points between 1999 and 2007, reaching 96%, above the Latin American average. In spite of this progress, only 44% of the Nicaraguan children who start primary school reach the last grade in six years. This is the lowest figure in the region, even compared to neighboring countries of similar level of development, such as El Salvador, Guatemala and Honduras. The challenge is ever more daunting in secondary education. The net enrolment rate in this level only reaches 46% in Nicaragua – the lowest rate in the region along with Guatemala’s – compared to 72% on average for Latin America. At the same time, almost 10% of Nicaraguan children between 5 and 14 years old are involved in work (IPEC, 2008). This figure rises to 14.4% among boys. Children who are economically active are less likely to attend school than their non-working peers (60% vs. 80%).

1528

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

Table 1 Net enrolment rates in primary and secondary education and survival to last grade of primary education in selected countries of Latin America. Net enrolment rates, primary education (%)

El Salvador Guatemala Honduras Nicaragua Latin America

Survival rate to last grade, primary education (%)

Net enrolment rates, secondary education (%)

1999

2007

2006

2007

– 82 – 76 93

92 95 93 96 94

69 62 81a 44 84

54 38b – 46 72

Source: UNESCO (2010). Notes: The figures for Latin America correspond to the weighted average for the net enrolments rates and the median for the survival rate. a Data are for the school year ending in 2005. b Data are for the school year ending in 2006.

A considerable amount of research on child labor and schooling decisions has been carried out in the past few years.1 However, these studies predominantly focus on the determinants of children’s work and school participation rather than on the causal impact of work on schooling. Furthermore, school enrollment is not a sufficient indicator of educational achievement, or of the quality of children’s school experience. Since many children combine both activities, the trade-off between work and school participation only provides us with indirect evidence of the potential consequences of child labor on human capital. Most of the empirical studies that attempt to estimate the direct consequences of child labor on educational attainments in developing countries have employed a static framework. The findings of these studies tend to support the conclusion that children’s work is detrimental for educational performance and attainments. For example, Psacharopoulos (1997) in urban Bolivia, Rosati and Rossi (2003) in Nicaragua and Pakistan, Ray and Lancaster (2005) in Belize, Cambodia, Namibia, Panama, Philippines and Portugal, Sedlacek, Duryea, Ilahi, and Sasaki (2005) in 16 Latin American countries, and Goulart and Bedi (2008) in Portugal found child labor to be significantly associated with school failure or low educational attainments. Other research focusing on cognitive achievements arrives at similar conclusions. This has been the case of Akabayashi and Psacharopoulos (1999) in Tanzania, Heady (2003) in Ghana, and Gunnarsson, Orazem, and Sánchez (2006) employing mathematics and language test scores from a sample of 3rd and 4th graders in 11 Latin American countries. On the other hand, a study by Sabia (2009) in the United States founds that much of the school-year employment effects on the academic performance of adolescents can be explained by individual heterogeneity. Even though there is a general consensus in the literature that children’s work is negatively associated with schooling outcomes, the static framework of analysis cannot assess the long-term consequences of household decisions about children’s work. Empirical research employing a dynamic framework to assess the consequences of child labor on human capital’s accumulation is limited, although the available studies reveal important

1 Edmonds (2007) provides a detailed review of the recent empirical literature.

patterns. Boozer and Suri (2001), employ data from different months of a single year in Ghana, and use regional rainfall patterns as the identification strategy for child labor intensities. They find that an hour of work is associated with a decline in school attendance of 0.38 h. Canals-Cerdá and Ridao-Cano (2004) make use of retrospective information to evaluate the impact of childhood work experiences on subsequent educational progress in rural Bangladesh. They find a negative and significant effect of child labor occurrence on school progress, with more damage the earlier in life a person begins to work. Beegle, Dehejia, and Gatti (2005) analyze the impact of children’s work on a series of schooling, health, and labor market outcomes, employing a panel data set from Vietnam and using an instrumental variables strategy, with community shocks and rice prices as instruments for child labor. They find that, at the mean level, children’s work leads to 30% lower chances of being in school and a 6% decrease in educational attainment five years later. Despite these negative effects, analyses of young adult wages reveal that the returns to experience exceed the returns to schooling. Similarly, Beegle, Dehejia, Gatti, and Krutikova (2008) study the impact of child labor on education, employment choices and marital status using a longitudinal survey in Tanzania. They find that a standard deviation increase in child labor hours is linked with a reduction of about half a year of schooling and with a decrease of over 8 percentage points in the probability of completing primary school. This paper improves on the empirical literature on the consequences of child labor on educational outcomes in a number of ways. First, it departs from most of the existing research, which has relied on static analyses, by examining outcomes over time. Second, it improves on the empirical research that employs a dynamic framework, by combining a series of features that are only partially present in these studies. Boozer and Suri (2001), for instance, measure the contemporaneous effect of hours of work on school attendance, which is not a very informative indicator of human capital stock. The empirical analyses carried out in this paper employ multiple schooling outcomes that are relevant for international commitments on education and measured in the medium term. Canals-Cerdá and Ridao-Cano (2004) employ work participation as their main explanatory variable, and use retrospective information, which may be affected by measurement error and memory bias. This paper makes use of time at work, shedding light on

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

a possible threshold from which child labor is particularly damaging. The approach in this paper has more similarities with the studies by Beegle et al. (2005, 2008), which analyze educational outcomes in the medium term, employ hours of work as the main explanatory variable, and estimate the model through instrumental variables. However, the authors make some restrictions in the sample, limiting the analysis to children who were enrolled in school during the first round of the survey or to those below age 10, thus leaving aside the time allocation decision in the first period. This limitation may introduce biases and reduces the generalization of results. This paper improves upon this approach by modelling children’s time allocations in the first period and including the previous human capital achieved as a regressor of present schooling outcomes, as a further indicator of past inputs. Also, Beegle et al. (2005) employ a more restrictive definition of work that excludes non-market production, while this paper includes those activities as they might also affect schooling outcomes. The rest of the paper is organized as follows: the data source and the main variables are described in Section 2, followed by the conceptual framework and empirical strategy in Section 3. The empirical results are discussed in Section 4, which also includes some specification tests. Section 5 concludes. 2. Data set and variables 2.1. Data set The empirical analysis is performed with panel data from the Nicaraguan Living Standards Measurement Survey (EMNV by its Spanish acronym) carried out in 1998 and 2001. The EMNV was developed with the objective of investigating the living conditions of households residing in the national territory of Nicaragua. In addition to a household, an anthropometrics and a price questionnaire, the 1998 EMNV included a time use module administered to all members of a household over 6 years old, containing detailed information of the activities performed during the previous day.2 The total number of households surveyed in 1998 was 4209. The EMNV 2001 was administered to all households in the 1998 sample that were found again within the limits of the updated census segment; households occupying dwellings which had been classified as ‘non-response’ in 19983 ; and a group of new dwellings and households added to reflect population growth in the period. The total number of households in the 2001 survey sample was 4676, but due to attrition between rounds the percentage of subjects from the original sample that could be re-interviewed was actually 74.5%. The sample of interest for this paper includes children who were between 6 and 14 years old in 1998, for who

2 The time use module was applied to one-half of the dwellings, but the selection was done randomly: administered to every other dwelling, changing the first dwelling in every other census segment. 3 A household was classified as “non-response” in 1998 if it refused to be interviewed, was inaccessible due to violence, was absent, the housing unit was uninhabited, or there were mapping errors.

1529

detailed time use information can be retrieved in that year, and in households which were re-interviewed in 2001.4 The cut-off ages were taken considering the minimum age for which the survey included time use information and the minimum age for admission to work established by ILO Convention No. 138, 15 years old. The EMNV 1998 sample included 5,976 children in this age group, of which 2946 have time use information. Attrition from 1998 to 2001 reduces the final sample size for the empirical analysis in this paper to 1955 children. Some analyses are performed to assess the potential biases due to sample attrition. 2.2. Description of variables Human capital accumulation is measured by four alternative variables: years of education, grade-for-age, completion of primary education, and completion of at least a year of secondary education. Years of education are measured by a discrete variable representing the years of schooling completed by a child by 2001. Grade-for-age is measured as the ratio between years of education attained by 2001 and the number of school-years that would be appropriate for a given age in Nicaragua.5 A ratio of one means that the child has attained the grade expected for his/her age, while a ratio below one means that the child is lagging behind with respect to his/her age. A ratio of zero means that the child has not acquired any schooling. Completion of primary education is measured by a binary variable indicating whether a child aged 13 years or more in 2001 has attained at least 6 years of schooling by 2001, conditional on not having achieved that level by 1998. Completion of at least a year of secondary school is also measured by a binary variable, indicating whether a child 14 years old or more has attained at least 7 years of education by 2001 conditional on not having achieved that level by 1998. 6,7 The explanatory variable of main interest is daily hours of work carried out by a child in 1998, a continuous measure censored at zero. Total work time is constructed by adding up the time information for all paid or unpaid work performed by a child for outside employers or for the family farm/business (market production activities), and all

4 In the case of children below age 15 years, the individual questions were answered by the household head or best informant in the household. 5 Grade-for-age = (completed years of education by 2001/age − 7). Seven was the official entry age to primary education in Nicaragua when the 1998 survey was taken. 6 The cut-off ages for the completion of primary school and completion of a year or more of secondary school outcomes are based on the ideal age range in primary education in Nicaragua: 7–12 years old. The outcome of completed primary school is computed for children who are 13 years old and more in 2001 and who had not completed primary education by 1998 (846 observations). The outcome of a year or more of secondary school is computed for children who are 14 years old and more in 2001and who had not completed a year of secondary education by 1998 (719 observations). 7 The schooling measures employed represent only some aspects of an individual’s human capital investment, even considering a narrower definition of human capital as educational achievements. For instance, the outcomes proposed are not indicators of cognitive attainment, as we would observe in learning test scores. However, given the data available, the outcomes proposed herein provide a good starting point to evaluate the effects of children’s work on human capital accumulation.

1530

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

Table 2 Summary statistics. Mean Dependent variables Years of education, 2001 Grade-for-age, 2001 Primary school complete, 2001 At least a year of secondary school, 2001 Variables to be instrumented Hours of work per day, 1998 Hours of work per day in market production, 1998 Hours of work per day in non-market production, 1998 Years of education, 1998 Grade-for-age, 1998 Independent variables Rural area Atlantic region Central region Pacific region Male Age, 1998 Adult hourly wage at municipal level, 1998 (ln) School fees per enrolled at municipal level, 1998 (ln) Teachers with diploma at municipal level, 1998 (%) Pupils/teacher in primary school at department level, 1998 Occurrence of nature shock, 1998 Distance to primary school, 1998 (meters) Access to a road, 1998 Electricity, 1998 Piped water, 1998 Value of durable goods, 1998 (ln) Years education household head, 1998 No. Household members, 1998 No. Children aged 0–5, 1998 Age, 2001 Adult hourly wage at municipal level, 2001 (ln) School fees per enrolled at municipal level, 2001 (ln) Teachers with post-secondary education at department level, 2001 (%) Pupils/teacher in primary school at department level, 2001 Occurrence of nature shock, 2001 Distance to primary school, 2001(meters) Access to a road, 2001 Electricity, 2001 Piped water, 2001 Value of durable goods, 2001 (ln) Years education household head, 2001 No. Household members, 2001 No. Children aged 0–5, 2001

Standard deviation

4.056 0.630 0.407 0.332

2.645 0.358 0.491 0.471

2.184 0.580 1.605 2.076 0.562

2.898 1.873 2.208 2.138 0.415

0.492 0.151 0.359 0.378 0.521 9.828 2.610 4.759 64.218 36.572 0.374 970.089 0.768 0.576 0.492 3.122 3.616 7.314 1.138 12.878 0.764 11.794 34.793 47.212 0.259 842.169 0.826 0.656 0.544 4.188 3.832 7.158 0.841

0.500 0.358 0.480 0.485 0.500 2.536 0.667 1.520 17.367 4.720 0.484 1843.310 0.422 0.494 0.500 3.705 3.775 2.775 1.154 2.555 1.159 0.844 10.907 23.839 0.438 1504.911 0.379 0.475 0.498 3.864 3.862 2.735 1.013

Notes: Statistics calculated using a panel sample of 1955 children 6–14 years old from the 1998 and 2001 Nicaraguan EMNV. The outcome of completed primary school is computed only for children who are 13 years old and more in 2001 and who had not completed primary education by 1998 (846 observations). The outcome of a year or more of secondary school is computed only for children who are 14 years old and more in 2001and who had not completed a year of secondary education by 1998 (719 observations). Price variables are in 1998 Córdobas and adjusted by a regional factor.

activities performed in household chores (non-market production activities) during the day previous to the survey. The empirical analysis is first performed with the aggregate measure of work, and afterward distinguishing between work in market and non-market production. Even though time use measures provide valuable information on children’s activities, the recall period employed by the time use module in the EMNV has some disadvantages. In particular, previous day measures may be subject to a high noise to signal ratio, as in the day of reference a number of children might not have carried out any work activity. This, in turn, could bias the results and weaken the association between work and school outcomes. On the other hand, an instrumental variable estimation procedure should correct biases due to random measurement error.

Table 2 presents summary statistics of all the variables employed. Table 3 shows children’s baseline participation rates in work activities. With the comprehensive definition of children’s work adopted in this paper, the proportion of children that perform some type of working activity during the day surpasses what is usually found in child labor statistics. As Table 3 shows, if only work in market production were considered 15% of the boys between 6 and 14 years old would be identified as workers as well as just over 6% of the girls. On the other hand, 61% of the girls and around 50% of the boys spend part of the day performing domestic activities. As a result, about three out of five children in the sample are engaged in some type of work activity, with girls participating in a slightly higher proportion than boys. Children’s work participation is more pronounced in the

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

1531

Table 3 Child labor participation rates by area and gender, 1998 (in percentages). Boys

Work in market production Household farm/business Outside work Total work in market production Work in non-market production Total work

Girls

Rural

Urban

Total

Rural

Urban

Total

14.1 8.9 22.7 55.8 65.3

2.0 6.6 8.5 47.6 51.9

7.5 7.6 15.0 51.4 58.0

3.8 3.7 7.5 64.9 66.7

1.0 4.5 5.5 58.4 59.5

2.3 4.2 6.4 61.3 62.8

Notes: Statistics calculated using a sample of 1955 children 6–14 years old from the 1998 Nicaraguan EMNV. Rates refer to participation in the activity the day previous to the survey. Totals may not add because children may perform more than one activity.

Table 4 Children’s participation and time allocation in work and schooling by age group, 1998. 6–9 years old

School Work in market production Work in non-market production Total work

10–14 years old

Total

Participation

Time in activity

Participation

Time in activity

Participation

Time in activity

64.4 5.0 45.1 47.1

6:16 4:13 2:25 2:43

64.1 16.2 65.8 71.9

6:32 5:21 2:59 3:58

64.2 10.9 56.1 60.2

6:25 5:06 2:46 3:30

Notes: Statistics calculated using a sample of 1955 children 6–14 years old from the 1998 Nicaraguan EMNV. Participation is measured in percentages, and time allocation is conditional on participating in the activity and measured in hours and minutes. A child can perform more than one activity per day.

rural areas, where it comprises two-thirds of the sample. Distinguishing between market and non-market production, 23% of the boys in rural areas are engaged in market work compared to about 9% of their counterparts in urban areas. The source of most of the differences in market work across geographical location and gender is the higher rate of engagement of rural boys in their family’s farms or businesses. In Table 4, children’s participation and time dedicated to work and schooling are broken down by age group. Work in market production increases with age: the share of children 6–9 years old performing this type of activities is 5%, but climbs to over 16% among children 10–14 years old. Time dedicated to the activity also increases with age, although the load for young children is still substantial. The youngest group dedicates near 4 h and a quarter in the day of reference, while the older group works an additional hour. Non-market production follows a similar pattern: younger children participate less in the activity and dedicate a lesser amount of time than older children. Still, 45% of children aged 6–9 years work in this type of activity, allocating nearly 2 and half hours to it in the day of reference. Two in three children 10–14 years old work in non-market production for almost three hours in the day of reference. Table 5 illustrates the relationship between school outcomes in 2001 and children’s past work experiences. The sample is separated in two groups defined by their age in 1998: 6–9 and 10–14, to account for higher levels of work participation as children grow up as well as different levels of education expected at each age group. Observing the first three rows under each age group in Table 5, which display educational outcomes by categories of activity, it is clear that a complete dedication to work is negatively associated

with children’s educational prospects in both age groups. In contrast, the combination of school and work is not necessarily associated with lower achievement, and in some cases is associated with better outcomes than attending school as a full-time activity. This is the case with mean years of education in both groups of children and with the proportion of older children that have completed primary school and at least a year of secondary school. A clearer picture emerges when hours of work are taken into account rather than participation. As it may be observed in the last four rows under each age group in Table 5, a small amount of work during a day does not appear to compromise schooling achievements in the medium term. Indeed, children who worked up to 1 h a day in 1998 had better schooling outcomes in 2001 than those who did not work at all. But above this level, educational attainments fall off increasingly, most notably over two and a half hours of daily work. Among children 10–14 years old, the rate of decrease in school achievements between the two upper levels of work intensity ranges from 29% to 58%.

3. Conceptual framework and empirical strategy To estimate the impact of child labor on educational outcomes, I employ a dynamic, unitary household decisionmaking model in the spirit of Baland and Robinson’s (2000) finite horizon model.8 In this model, a child is endowed with one unit of time in each period, which can be allocated

8

A fully developed model is available as supplementary material online.

1532

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

Table 5 School outcomes in 2001 by work participation and school attendance in 1998 and by quartiles of time at work in 1998. School outcomes in 2001

Ages 6–9 School only Work and school Work only Work less than 1 h Work 1–2.5 h Work 2.5–5 h Work more than 5 hours Ages 10–14 School only Work and school Work only Work less than 1 h Work 1–2.5 h Work 2.5–5 h Work more than 5 hs

Years of education

Grade-for-age

Primary school complete

At least a year of secondary school

3.306 3.545 1.835 3.541 2.808 2.423 2.231

0.746 0.702 0.477 0.745 0.604 0.538 0.516

– – – – – – –

– – – – – – –

5.817 6.214 3.925 6.176 5.702 5.665 3.697

0.733 0.719 0.480 0.794 0.667 0.660 0.431

0.464 0.543 0.265 0.593 0.471 0.468 0.223

0.446 0.448 0.186 0.591 0.401 0.368 0.156

Notes: Statistics calculated using a panel sample of 1955 children 6–14 years old from the 1998 and 2001 Nicaraguan EMNV. The outcome of completed primary school is computed only for children who are 13 years old and more in 2001 and who had not completed primary education by 1998 (846 observations). The outcome of a year or more of secondary school is computed only for children who are 14 years old and more in 2001 and who had not completed a year of secondary education by 1998 (719 observations). Due to the age sample restrictions to compute primary school and secondary school outcomes, only values for children over 10 years old are shown in the table for these two outcomes. Categories of time at work are conditional on work participation.

to work in the production of the household output (th m ) and/or to the production of human capital (ts m ).9 h s 1 = tm + tm

for m = 1, 2

(1)

Human capital is produced through an increasing and concave function of the time spent by the child in schooling in the same period, the human capital previously accumulated (Sm − 1 ), and exogenous school quality and availability (Am ).10 s Sm = fm (tm ; Am ) + Sm−1

for m = 1, 2

(2)

The optimal time for schooling activities in the first two periods (ts 1 and ts 2 ) is found substituting the time, technology and budget constraints into the utility specification. Substituting the optimal school time allocation in the human capital production function for each period yields: S1 = f1 (t1s ∗; A1 ) + S0 S2 =

f2 (t2s ∗; A2 ) + S1

(3) =

f2 (t2s ∗; A2 ) + f1 (t1s ∗; A1 ) + S0

(4)

Through the time constraint we have that: t1s ∗ = 1 − t1h ∗ Thus, the human capital accumulated in period two can also be expressed: S2 = f2 (1 − t2h ∗; A2 ) + f1 (1 − t1h ∗; A1 ) + S0

(5)

9 Only about 2.5% of the children in the sample were found to be working for a wage; instead most of the work they performed was unpaid and related to household-based production (e.g. agricultural and small business) or household chores. 10 Availability refers to the supply of schools, or the school grades relevant for the child’s schooling decision, or any other measure of access to educational opportunities.

Accordingly, the level of human capital acquired by a child when she is, for example, of secondary schooling age will be a function of the household’s decisions on her time allocation in that same period, but also in the previous period (besides own endowment of human capital). Therefore, an empirical analysis of school outcomes should take into account the endogeneity of the time allocation decision in order to obtain more precise estimates. The impact of children’s work on schooling outcomes can be estimated through a linear regression approximating function (5), which expressed educational outcomes as depending on inputs allocated in past periods in addition to inputs in the present. One of these inputs is time available for schooling activities, or, through the time constraint, time allocated to work. Therefore, the impact of children’s work could be estimated as the coefficient on children’s time at work in a linear regression. A linearized version of the education production function in (5) could take the following form: Si,t = ˇ0,t + ˇ1 Wi,t−3 + ˇ2 Si,t−3 + ˇ3 Zi,t + εi,t

(6)

where Si,t is the educational attainment of child i measured at the beginning of the present period, Wi,t − 3 is the hours worked in a day by child i three years earlier, Si,t − 3 is the educational attainment of child i three years earlier, and Z i,t are controls for individual and household characteristics, school quality and availability, location, etc. The error term, εi,t captures time-varying unobservable characteristics of child i that might affect present schooling outcomes and that are unknown at time t. To obtain consistent estimates of the ˇs with OLS, the error should have zero mean and be uncorrelated with the regressors. Note that this function does not include time at work in the present period as one of the explanatory variables. The main reason is that the schooling outcomes measured in

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

the survey represent the stock of education acquired by the child by the beginning of the present period, that is, is the result of past inputs rather than present ones.11 Nevertheless, the linear regression will include as controls measures which affect children’s time allocation in the present. Ideally we would have information on all personal, family, community, and school inputs in the past history of the child, but in this case we only observe two periods. Accordingly, the production function includes two measures of past inputs: time allocated to work three years earlier, and schooling achieved by the beginning of the period t − 3, assuming the latter is a sufficient indicator of prior inputs. This formulation allows examining changes in school progress between two periods, with a child’s labor experience in the previous period as the main factor of interest explaining this change. An additional concern relates to the interpretation of the results from the dynamic specification in this study. Following the specification in (6), prior schooling would also be affected by prior time at work, and also the impact of work at a point in time would be forwarded to subsequent periods. What will be actually estimated in this paper is a single-period effect of child labor intensity on subsequent school outcomes, not the complete dynamic system. Accordingly, the estimation will not be addressing the full dynamic implications of the model, and the estimate of prior time at work will not be measuring the total impact of children’s work on their human capital accumulation. Rather, the estimation will offer a more limited assessment of the impact of child labor on school progress in the period covered by the surveys. The inclusion of lagged inputs has some additional drawbacks. The human capital production function may have a time-persistent component, in which case the error term εi,t will be correlated with the lagged schooling measure. For example, a child’s ability or unobserved characteristics of the family may influence both work and schooling decisions causing the error terms to be correlated with the schooling measure. Moreover, children’s time in work and school activities is the result of household decisions making intertemporal trade-offs in children’s time allocation to maximize utility, making time at work and school potentially endogenous. I attempt to address potential biases stemming from correlated disturbances and endogeneity employing instrumental variables. The estimation strategy involves two steps that may be summarized as follows: first, I estimate two equations, one for each of the potentially endogenous prior work and schooling variables, as functions of a set of instruments; then, the predicted values based on these equations are introduced in place of the prior work and schooling variables in the estimation of school outcomes, Eq. (6). The specifications of the first stage regressions are: Wi,t−3 = 0,t−3 + 1 Bi,t−3 + 2 Zi,t + ωi,t−3

(7)

Si,t−3 = ı0,t−3 + ı1 Bi,t−3 + ı2 Zi,t + ωi,t−3

(8)

11 This would have been different if we had been measuring school enrollment in the present period.

1533

where Wi,t − 3 , Si,t − 3 and Zi,t are defined as before; Bi,t − 3 is a vector of instruments for time allocation of child i measured in t − 3; and ωi,t − 3 represents unobserved characteristics of child i in period t − 3. The second stage equation incorporates the predicted values of past time at work and school outcome obtained in the first stage, in the following manner: ˆ i,t−3 + ˇ2 Sˆ i,t−3 + ˇ3 Zi,t + εi,t Si,t = ˇ0,t + ˇ1 W

(9)

ˆ i,t−3 is the predicted expected time at work for where W child i in period t − 3, and Sˆ i,t−3 is the predicted school outcome for child i in period t − 3. The set of variables introduced to provide exogenous shifts in the child labor and school time allocation equations includes the following: whether a household was affected by a natural shock in the past 12 months, adult hourly wage at the municipal level, average school fees per enrolled at the municipal level, percentage of teachers with diplomas at the municipal level, pupil/teacher ratios in primary school at the department level, distance to primary school, household access to electricity, piped water, road, value of household durable goods, years of education of the household head, number of children under five years old and household size. The first stage regressions also control for sex and age of the child, urban/rural residence and geographical regions. In order to reduce the possibility that the selected instruments may be correlated with unobserved factors affecting present school achievements, thus affecting these outcomes directly, I include as controls in the present school outcome equations similar measures to those employed to model time allocation in the past, but measured in the present (i.e. adult hourly wages at community level, measures of school prices, quality and availability at community level, presence of natural shocks affecting production, value of household durable goods, household access to electricity, piped water and roads, and other household factors measured in the present). The two equations in the first stage of the instrumental variables strategy are estimated by different methods, as the endogenous variables differ in their measurement. Time allocated to work is measured in hours and, given that is a censored outcome, is estimated by maximum likelihood. The fitted value to be incorporated in the second stage equation corresponds to the expected value of time at work from the maximum likelihood estimation. For the case of school achievement in the past, two alternative measures are constructed: years of education, and a modification of it, grade achieved relative to the age of the child. In both cases, school achievement in the first stage is estimated by ordinary least squares, and its fitted value introduced in the respective second stage regression. The estimation method for the second stage regression varies according to the type of schooling outcome that we are interested in: ordinary least squares is employed for interval measures of school outcomes and probit is employed for binary measures of school outcomes. Standard errors in the second stage are obtained through bootstrapping. A limitation of the instrumental variables estimation approach is that weak or inappropriate instruments may

1534

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

bias the results. In order to assess the suitability of the approach and the employed specification I conduct a series of tests: first I test the null hypothesis that the suspected regressors are exogenous, then I test the joint significance of the instruments, to assess their strength, and finally, I test the joint validity of the exclusion restrictions. A final concern for the validity of the results is that sample attrition resulting from the longitudinal nature of the data may introduce biases. For this reason, a series of tests are conducted to evaluate if attrition poses a threat to the validity of the findings. 4. Results 4.1. Effect of total time at work Estimation results from the second-stage regressions for the five schooling outcomes are presented in Table 6, as well as OLS and probit estimation results employing the potentially endogenous regressors directly.12 The amount of time a child dedicated to work in the past has a negative impact on schooling outcomes in all but one of the estimations performed. Furthermore, the marginal effects obtained by two-stage procedures are larger than those obtained with direct measures of working time and schooling. This suggests that directly introducing past work experience in the schooling equations underestimates the detrimental effect of children’s work. The first row and pair of columns of Table 6 display the effect of child labor on the years of education attained by a child by 2001. The estimates indicate that an increase of one hour in work time in 1998 is associated with a reduction of almost 0.4 years of schooling completed by 2001, or over 4 months, controlling for the level of education already achieved and other exogenous factors. In contrast, if past time at work and years of schooling are taken as predetermined in an OLS regression, the impact of time at work is smaller: an hour increase of time at work results in a gap in school years of less than a month. Previous work experienced by children is also significantly and negatively related to the grade-for-age achieved by 2001. An increase of one hour in daily work carried out in 1998 leads to a deterioration of the grade-for-age measure of 2 percentage points three years later. For a child at the mean value of this indicator in 2001, which at 0.63 already represents a considerable delay, the change brought about by increasing child labor by 1 h would represent an additional lag of 3%. Although the coefficient of time at work in the 2SLS is almost twice the obtained with OLS, in practice the marginal effects are not very different if one computes the changes in standard deviations from the mean in each case. The impact of children’s work on the probability of primary school completion is presented in the third pair of columns of Table 6. Although the impact of past hours of work in the two-stage probit is insignificant, the effect is statistically different from zero at the 5% level when time

12 For space reasons the results of the first-stage regressions are omitted in this paper, but are available upon request.

at work is included directly as an exogenous covariate in the estimation of the probability of primary school completion. In the latter specification, an hour increase in daily time at work in 1998 is linked to a fall in the probability of completing primary school of 2 percentage points three years later. The last pair of columns in Table 6 presents the marginal effects on the probability of completion of a year or more of secondary school. In contrast to the other school outcomes, the likelihood of making the transition to secondary education between the survey rounds increases with daily hours dedicated to work in 1998 using the instrumental variables. On the other hand, the probit model including time at work and years of education in the past period as exogenous variables yields the expected negative estimate on time at work, but it is not significantly different from zero. Along with the marginal effects of children’s time at work, Table 6 summarizes the marginal effects of other covariates included in each model. Among these variables are indicators of previous human capital accumulation. Lagged years of education and lagged grade-for-age have a statistically significant effect in all specifications. Taking the probability of primary school completion, for example, an additional year of education attained by 1998 leads to an increase of 0.31–0.35 percentage points in the likelihood of completion of elementary school, depending on the specification. Some of the other explanatory variables have significant coefficients in the school outcome models, but this varies across specifications. In particular, boys exhibit worse school outcomes than girls in most models, although the differences are not always statistically significant. This finding is consistent with the national trends, where despite gender parity in primary school enrollment rates, girls tend to perform better in school. Additionally, in all models but the 2SLS specification of years of education, the age of the child is also significantly associated with lower school progress. This could reflect the increasing opportunity costs of schooling as children get older, especially if they are boys, or it could reflect cohort effects, as older children might have experienced less educational opportunities than their younger peers. 4.2. Levels of work effort The next set of estimates assesses whether different levels of work effort experienced in the past have dissimilar impact on a child’s educational achievements. For this purpose, I re-estimate the second-stage models corresponding to the five school outcomes introducing a piecewise linear spline with knots at the 50th and 75th percentiles of the expected time at work predicted in the first stage regression.13

13 A more rigorous procedure would be to estimate first-stage regressions with the spline segments as dependent variables. I followed a more informal procedure in order to simplify the interpretation and streamline the presentation, so the results should be taken as rough estimates of the impact of work intensity levels on school outcomes.

Table 6 Marginal effects of the determinants of school outcomes in 2001. School outcomes in 2001 Years of education

Completion of one year or more of Secondary school

2SLS

OLS

2SLS

OLS

2SPr.

Probit

2SPr.

Probit

−0.385*** (0.071) 1.028*** (0.152) – 0.025 (0.237) −0.792 (0.587) −0.774** (0.375) −0.501 (0.372) −0.317*** (0.105) 0.105 (0.096) −0.071* (0.040)

−0.041*** (0.009) 1.104*** (0.019) – −0.184 (0.116) −0.421 (0.298) −0.507*** (0.196) −0.209 (0.190) −0.201*** (0.045) −0.070*** (0.015) −0.004 (0.023)

−0.020*** (0.006) – 0.663*** (0.115) −0.045 (0.042) −0.049 (0.097) −0.060 (0.061) −0.026 (0.063) −0.014 (0.017) – −0.009 (0.007)

−0.012*** (0.002) – 0.578*** (0.020) −0.033 (0.029) −0.009 (0.071) −0.017 (0.045) 0.019 (0.046) −0.030*** (0.011) – −0.014** (0.006)

−0.029 (0.092) 0.305*** (0.205) – −0.024 (0.266) −0.010 (0.705) 0.027 (0.461) 0.034 (0.466) −0.063 (0.143) −0.096* (0.137) −0.024 (0.054)

−0.019** (0.030) 0.352*** (0.109) – 0.088 (0.358) −0.202 (0.973) −0.238 (0.647) −0.169 (0.614) −0.095** (0.151) −0.046*** (0.060) −0.004 (0.073)

0.072* (0.121) 0.323*** (0.264) – −0.082 (0.310) 0.252 (0.800) 0.276 (0.547) 0.219 (0.523) 0.031 (0.166) −0.160*** (0.183) −0.010 (0.082)

−0.002 (0.033) 0.100*** (0.088) – −0.013 (0.354) 0.023 (0.999) 0.048 (0.724) 0.035 (0.671) −0.003 (0.166) −0.020*** (0.085) −0.010 (0.106)

−0.247** (0.108)

−0.210*** (0.061)

−0.035* (0.019)

−0.002 (0.014)

−0.017 (0.138)

−0.049 (0.196)

0.084 (0.182)

0.017 (0.256)

−0.002 (0.015)

0.008 (0.008)

0.002 (0.003)

0.001 (0.002)

−0.002 (0.018)

−0.004 (0.024)

−0.001 (0.020)

−0.001 (0.023)

0.003 (0.003)

0.002 (0.001)

4.e−4 (5.e−4)

4.e−4 (3.e−4)

2.e−4 (0.003)

0.001 (0.004)

3.e−5 (0.005)

−1.e−5 (0.006)

0.197 (0.115)

0.121 (0.063)

0.008 (0.021)

0.019 (0.014)

0.049 (0.164)

0.053 (0.222)

−0.049 (0.223)

−0.025 (0.260)

1.e−5 (3.e−5)

−5.e−5*** (2.e−5)

−7.e−6 (4.e−6)

−6.e−6** (2.e−6)

1.e−5 (5.e−5)

2.e−5 (5.e−5)

2.e−5 (6.e−5)

1.e−5 (9.e−5)

−0.053 (0.121) 0.149 (0.144) −0.028 (0.116) 0.001 (0.018)

−0.034 (0.071) 0.207*** (0.076) 0.048 (0.061) 0.031*** (0.008)

0.013 (0.021) 0.029 (0.028) −0.003 (0.020) 0.003 (0.003)

0.013 (0.016) 0.032* (0.017) 0.001 (0.014) 0.005** (0.002)

−0.074 (0.169) 0.068 (0.180) −0.023 (0.143) −1.e−4 (0.024)

−0.009 (0.240) 0.057* (0.239) 0.016 (0.187) 0.007 (0.023)

−0.004 (0.250) 0.181** (0.251) −0.022 (0.181) 0.011 (0.030)

0.021 (0.351) 0.073*** (0.346) −0.010 (0.247) 0.005** (0.028)

0.027* (0.015)

0.028*** (0.007)

0.002 (0.002)

−2.e−5 (0.002)

0.009 (0.023)

0.018*** (0.024)

−0.004 (0.026)

0.003 (0.025)

−0.038* (0.022)

−0.025** (0.011)

0.001 (0.004)

−0.002 (0.003)

0.001 (0.030)

2.e−4 (0.035)

−0.001 (0.039)

−0.005 (0.038)

0.031 (0.055)

−0.014 (0.031)

−0.006 (0.010)

−0.004 (0.007)

−0.043 (0.072)

−0.034 (0.092)

−0.029 (0.105)

0.006 (0.132)

0.575 – 1955

0.865 – 1955

0.251 – 1955

0.626 – 1621

– −420.629 846

– −187.473 846

– −301.830 719

– −150.824 719

*

*

*

1535

Notes: Estimates of the probability of completed primary school are computed only for children who are 13 years old and more in 2001 and who had not completed primary education by 1998. Estimates of the probability of a year or more of secondary school is computed only for children who are 14 years old and more in 2001and who had not completed a year of secondary education by 1998. Marginal effects in probit models are evaluated at the mean sample value of each variable. For dummy variables, the marginal effect is the discrete change from 0 to 1. Standard errors of the coefficients appear in parentheses; they are bootstrapped (1000 replications) for two-stage regressions, and robust for OLS and Probit regressions. * Significant at 0.10 level. ** Significant at 0.05 level. *** Significant at 0.01 level.

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

Hours of work in 1998 Years of education in 1998 Grade-for-age in 1998 Rural Atlantic Central Pacific Male Age, 2001 Adult hourly wage mun. level, 2001 (ln) School fees mun. level, 2001 (ln) Teachers with post-secondary education dept. level, 2001 (%) Pupils/teacher in primary school dept. level, 2001 Production affected by nature shock, 2001 Distance to primary school, 2001 (meters) Access to a road, 2001 Electricity, 2001 Piped water, 2001 Value of durable goods, 2001 (ln) Years education household head, 2001 No. household members, 2001 No. children aged 0–5, 2001 R2 Log likelihood N

Completion of Primary school

Grade-for-age

1536

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

Table 7 Marginal effects of segments of hours worked in 1998 on two-stage estimations of school outcomes in 2001. School outcomes in 2001

Segments of the spline Worked below the median (up to 1:55 h) Worked between 50th and 75th percentiles (1:55–3 h) Worked more than the 75th percentile (3 h +) N

Years of education

Grade-for-age

0.828*** (0.154) −0.155 (0.141) −0.346*** (0.087) 1955

0.019 (0.020) 0.056 (0.272) −0.035* (0.021) 0.065 (0.198) −0.032*** (0.010) −0.054 (0.110) 1955 846

Completion of primary school

Completion of one year or more of secondary school 0.107 (0.396) 0.135* (0.242) 0.054 (0.140) 719

Notes: Estimates of the probability of completed primary school are computed only for children who are 13 years old and more in 2001 and who had not completed primary education by 1998. Estimates of the probability of a year or more of secondary school is computed only for children who are 14 years old and more in 2001and who had not completed a year of secondary education by 1998. Marginal effects correspond to two-stage regressions, and in the case of the probit models are evaluated at the mean sample value of each segment of the spline. Standard errors of the coefficients appear in parentheses and are bootstrapped (1000 replications). Other controls included in the regression but omitted from the table are: predicted years of education in 1998, sex, area of residence, and other controls measured in 2001, and listed in Table 1. * Significant at 0.10 level. *** Significant at 0.01 level.

Table 7 presents marginal effects and standard errors corresponding to the slope estimates of each segment in the spline. The first segment measures the impact of past child labor on subsequent school achievements if the predicted time at work is below the median. The second segment measures the impact when the predicted time at work in 1998 is between the median and the 75th percentile. Lastly, the third segment measures the impact when the predicted time at work is over the 75th percentile. As may be observed in Table 7, there is a certain degree of non-linearity in the effect of previous work experience on subsequent school outcomes. Working below the median predicted time – almost two hours a day – has, in fact, positive effects on all of the educational outcomes measured. In the case of completed years of education, for instance, a child for whom the predicted time at work in 1998 was less than two hours is expected to increase her/his educational attainment in 2001 by about 10 months for each additional unit of working time. For the other school outcomes, however, the estimates of the first segment of time at work are insignificant at the 10% level. At higher levels of work intensity the impact of past child labor experience on school achievements diverges depending on the outcome measure considered. For completed years of education and grade-for-age, the impact of child labor supply over the median level of two hours is negative. On the other hand, the probabilities of completing primary school or over a year of secondary school, increase for each additional unit of time worked between 2 and 3 h. The cost in terms of educational outcomes is evident for time worked over the 75th percentile in all but one of the schooling measures assessed. Each additional hour worked over three hours a day is associated with an extra loss of about 4 months of educational attainment three years later. The delay in school progress is also pronounced when we consider work over the 75th percentile: for each additional hour at this level of effort, children experience a deterioration of their grade for age measure of 3 percentage points. Additionally, the probability of completing primary school by 2001 is about 5 percentage points lower for each additional hour in children’s past work if they did so for

more than three hours, although the impact is not statistically significant at the 10% level. The findings described above would confirm that while child labor and schooling may be compatible activities, there is a threshold above which the effort has harmful consequences for school participation as well as achievements. 4.3. Work in market production or household chores In order to identify the existence of different impacts on educational achievements by type of work performed in the past, the school outcomes models are re-estimated creating two sets of specifications: one for work in market production and the second for work in non-market production – or household chores. Marginal effects and standard errors of hours of work from two-stage regressions, as well as estimates from standard OLS and maximum likelihood models are presented in Table 8. Time dedicated to work in any of the two activities has detrimental effects on most school outcomes, with the exception of the two-stage probit estimates for the completion of a year or more of secondary education. The effect of both work variables is statistically significant at least at the 10% level in most specifications. The amount of work in market production carried out by a child in the past appears to be more damaging than the amount of household work in the models of the grade for age attained and the probability of completing primary education. For example, an extra hour of work in market production in 1998 is associated with a worsening in school progress three years later, measured by the grade-for-age, which is between 48% and 75% larger than the marginal effect of the work performed in domestic chores, depending on the estimation procedure. In the model of years of education attained, the magnitude of the impact differs with the estimation: the size effect of market work is higher than that of household chores in the OLS model but lower in the two-stage model. In the former model, an additional hour of work in market production activities leads to a loss of 6 months in years of education attained three years later, whereas having

−0.002 (0.041) 0.113* (0.209) −0.011 (0.152) −0.009*** (0.002) −0.027*** (0.011) −0.025** (0.012) −0.500*** (0.124)

Notes: Statistics calculated using a panel sample of 1955 children 6–14 years old from the 1998 and 2001 Nicaraguan EMNV. The outcome of completed primary school is computed only for children who are 13 years old and more in 2001 and who had not completed primary education by 1998 (846 observations). The outcome of a year or more of secondary school is computed only for children who are 14 years old and more in 2001and who had not completed a year of secondary education by 1998 (719 observations). Each row represents a separate regression. In the case of the probit models, the marginal effects are evaluated at the mean sample value. Other controls included in the regression but omitted from the table are: predicted years of education in 1998, sex, area of residence, and other controls measured in 2001, and listed in Table 1. Standard errors of the coefficients appear in parentheses; they are bootstrapped (1000 replications) for two-stage regressions, and robust for OLS and Probit regressions. * Significant at 0.10 level. ** Significant at 0.05 level. *** Significant at 0.01 level.

−0.002 (0.043) 0.054* (0.101)

−0.010 (0.032)

Probit 2SPr. Probit

−0.026** (0.050) −0.060* (0.089)

2SPr. OLS

−0.014*** (0.002) −0.467*** (0.079)

2SLS OLS

−0.050*** (0.014)

2SLS

−0.040*** (0.009)

Completion of Primary school Grade-for-age Years of education

School outcomes in 2001

Table 8 Marginal effects of hours worked in 1998 from regressions of school outcomes in 2001 by type of work.

Hours of work in market production, 1998 Hours of work in non market production, 1998

Completion of one year or more of Secondary school

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

1537

worked in non-market production has half this effect. In the two-stage model an additional hour dedicated to either type of activity lowers the level of education attained in about 6 months. Employing OLS leads to an underestimation of the impact of working in non-market production on years of education. 4.4. Specification tests 4.4.1. Endogeneity of past time at work and school attainment The Wu–Hausman test was employed to assess the endogeneity of lagged measures of time allocation in the estimation of years of education and grade-for-age, and the Smith–Blundell test was used for the estimations of primary school completion and completing a year of secondary school.14 The null hypothesis that prior work and schooling are exogenous is rejected in the estimation of years of education and grade-for-age (p = .012 for years of education and p < .001 for grade-for-age). Therefore, the instrumental variables approach seems to be appropriate for these outcomes. Conversely, we cannot reject the exogeneity of past time at work and years of education in the other two school outcome models (p = .272 for primary school completion, and p = .619 for a year of secondary school). Accordingly, in the models for the probability of primary school completion and completion of at least a year of secondary education, the maximum likelihood estimates would be still consistent if we included time at work and years of education in 1998 directly as explanatory variables. However, as seen earlier, the effect of hours of work is statistically significant only in the probit model for primary school completion. As there are more instruments in each of the firststage regressions than potentially endogenous regressors I conducted tests of overidentifying restrictions. The null hypothesis in Sargan’s test – that the excluded instruments are jointly uncorrelated with the error term of the second stage regression – cannot be rejected in two of the school outcome models at conventional significance levels (p = .776 for primary school completion and p = .734 for a year of secondary education). The exceptions are the models of years of education attained by 2001 and the grade-for-age in 2001 (p = .002 for the former and p = .017 for the latter). As a consequence, the estimates of the years of education and the grade-for-age in 2001 by two-stage procedures may be biased, thus the inferences that were made about these two models should be qualified. A final test was conducted to assess the strength of the instruments in identifying past time at work and school attainment. In all the first-stage regressions, the instruments are jointly significant at the 1% level employing the F statistic, and therefore possess large explanatory power of time at work, years of education, and grade-for-age in 1998.15

14 The results of specification tests are available as supplementary material online. 15 Sensitivity checks were also conducted to assess the appropriateness of including household composition variables as instruments of past child

1538

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

4.4.2. Attrition in the sample The validity of the empirical analyses carried out in the previous sections could be compromised by the loss of sample observations between the two rounds of the EMNV survey if attrition were systematically related to the outcomes of interest. About one-fourth of the households in the original 1998 EMNV survey were lost between rounds due to migration, refusal to participate, or other reasons. Attrition is even larger for the subpopulation under study in this research, reaching about one-third of the original sample of children 6–14 years old who answered the time use questionnaire. In order to assess whether attrition biases the estimates, I conducted a series of tests based on Becketti, Gould, Lillard, and Welch (1988), and Fitzgerald, Gottschalk, and Moffit (1998).16 In very broad terms, these tests consist of: assessing whether the outcome variables of interest – in this case the four school outcomes – measured in the initial survey round are significant predictors in a binary choice model of attrition; assessing whether children leaving the sample between rounds exhibit different behavioral relationships than those that are re-interviewed, in terms of the determinants of schooling outcomes measured in the initial survey round; and assessing whether children that are followed-up in 2001 behave differently than those in the original sample in terms of the determinants of schooling outcomes measured in 1998. Overall, attrition in the sample under study does not appear to be causing major biases in the estimates of schooling outcomes. The tests lead me to reject the hypothesis that attrition bias is present in all the analysis of schooling outcomes. Therefore, the estimates obtained can be considered consistent. 5. Conclusion The main goal of this paper has been to assess the causal effects of child labor on subsequent educational achievements in Nicaragua. For this purpose, a series of school outcome models were constructed where the main explanatory variable is the time a child spent working during the reference day three years before. This paper improves upon previous research by combining a series of features only partially present in those studies: it employs panel data to assess the effects of child labor over time, centers on hours of work instead of participation – to identify differences in outcomes according to the intensity of work, includes household chores as well as work in market production, models children’s allocation of time in the first period, and includes the previous human capital achieved

labor and years of education attained, given the potential endogeneity of these variables in the estimation of child labor. To this end, all IV models were re-estimated without including two variables as covariates in either stage: household size and number of young children. The re-estimations do not differ in essence from the original ones. Neither the tests of exogeneity of child labor and years of education in 1998, nor the tests of the validity of the instruments present variations with respect to what was found for the complete models. 16 Falaris and Peters (1998), and Falaris (2003) derive a similar test as that developed by Becketti et al. (1988).

as a regressor of present schooling outcomes, as a further indicator of past inputs. Estimations were performed in two stages, including in the second stage the predicted time at work in 1998 and the predicted school attainment in 1998, in order to reduce biases due to potential endogeneity of these variables on school outcome equations in 2001. The estimations confirm previous findings that the time a child dedicated to work in the past has a harmful effect on most of her/his subsequent educational achievements. In terms of magnitude, an increase of one standard deviation over the mean daily hours of work carried out in 1998 is associated with a reduction of 27% on a child’s maximum education attainment, and with an additional delay in school progress of 9%. In other words, the trade-off with time for schooling activities implied by working for a longer amount of time accumulates to school failure in the medium term, either reflected in repetition or early school drop out. As a consequence, working children would have attained fewer years of education as they enter adulthood, thus damaging their potential earnings ability in the future. These results are in line with those obtained for Vietnam by Beegle et al. (2005). Understanding how children’s work experience affects access and completion of primary school is especially important in the context of the educational goals committed by the international community. During a similar period to the one considered in this research, 1998–2002, Nicaragua experienced an increase in the net enrollment ratio of primary education of almost 10% (UNESCO, 2005). However, during this period of expansion the school attainments of some children might have contracted in part due to their engagement in work activities. According to the findings in our research, the likelihood of a child completing primary school is negatively affected by her/his past work experience. A standard deviation increase in child labor in 1998 leads to a reduction in the probability of completing primary school of over 20 percentage points measured three years later. This suggests that in order to reach the universal primary education goal by 2015, Nicaragua would need to strengthen support to strategic groups, among them the group of children that are trailing behind due to their work effort. It has been argued that the educational threshold to achieve a decent standard of living in Latin America is complete secondary education (ECLAC, 2002); but secondary school enrollment rates in Nicaragua are among the lowest in Latin America. Among children of the relevant age who worked in 1998, the rates of completion of at least a year of secondary education by 2001 range from 19% to 45%, decreasing with the number of hours worked. However, when we control for other factors, the estimated effect of an additional hour of work is positive in the two-stage specification and negative, but not statistically significant in the probit specification. The intensity of work has also repercussions on a child’s schooling outcomes. For one part, having worked for up to two hours a day in 1998 is positively associated with all the educational outcomes measured. Although a specific threshold cannot be derived from the empirical analyses, the results suggest that working over three hours a day is

M. Buonomo Zabaleta / Economics of Education Review 30 (2011) 1527–1539

associated with the largest deterioration on a child’s school progress. In order to assess whether the type of work carried out in the first round of the survey affect the subsequent school outcomes in a different manner, the two-stage models were replicated for work in market production and in non-market production, separately. Although both types of work have negative effects on the majority of the educational outcomes measured in the medium term, time spent in market production activities in 1998 appears to be more damaging than time devoted to household chores, especially on grade-for-age and the probability of completing primary school. Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.econedurev.2011.08.008. References Akabayashi, H., & Psacharopoulos, G. (1999). The trade-off between child labor and human capital formation: A Tanzanian case study. Journal of Development Studies, 35(5), 120–140. Baland, J. M., & Robinson, J. A. (2000). Is child labor inefficient? Journal of Political Economy, 108(4), 663–679. Becketti, S., Gould, W., Lillard, L., & Welch, F. (1988). The panel study of income dynamics after fourteen years: An evaluation. Journal of Labor Economics, 6(4), 472–492. Beegle, K., Dehejia, R., & Gatti, R. (2005). Why should we care about child labor? The education, labor market, and health consequences of child labor. World Bank Policy Research Working Paper, 347, 9. Beegle, K., Dehejia, R., Gatti, R., & Krutikova, S. (2008). The consequences of child labor: Evidence from longitudinal data in rural Tanzania. World Bank Policy Research Working Paper, 467, 7.

1539

Boozer, M. A., & Suri, T. K. (2001). Child labor and schooling decisions in Ghana. Working paper, Yale University. Canals-Cerdá, J., & Ridao-Cano, C. (2004). The dynamics of school and work in rural Bangladesh. World Bank Policy Research Working Paper, 333, 0. Economic Commission for Latin America and the Caribbean (ECLAC). (2002). Social panorama of Latin America 2001–2002. LC/G.2183-P/I. Santiago, Chile. Edmonds, E.V. (2007). Child labor. IZA Discussion Paper 2606, Bonn. Falaris, E. M. (2003). The effect of survey attrition in longitudinal surveys: Evidence from Peru, Cote d’Ivoire, and Vietnam. Journal of Development Economics, 70(1), 133–157. Falaris, E. M., & Peters, H. E. (1998). Survey attrition and schooling choices. Journal of Human Resources, 33(2), 531–554. Fitzgerald, J., Gottschalk, P., & Moffit, R. (1998). An analysis of sample attrition in panel data: The Michigan panel study of income dynamics. Journal of Human Resources, 33(2), 251–299. Goulart, P., & Bedi, A. S. (2008). Child labour and educational success in Portugal. Economics of Education Review, 27(15), 575–587. Gunnarsson, V., Orazem, P., & Sánchez, M. A. (2006). Child labor and school achievement in Latin America. The World Bank Economic Review, 20(1), 31–54. Heady, C. (2003). The effect of child labor on learning achievement. World Development, 31(2), 385–398. International Programme on the Elimination of Child Labour (IPEC)-ILO. (2008). Nicaragua: Child labour data country brief. ILO: Geneva. Psacharopoulos, G. (1997). Child labor versus educational attainment: Some evidence from Latin America. Journal of Population Economics, 10(4), 377–386. Ray, R., & Lancaster, G. (2005). The impact of children’s work on schooling: Multicountry evidence. International Labour Review, 144(2), 189–210. Rosati, F. C., & Rossi, M. (2003). Children’s working hours and school enrolment: Evidence from Pakistan and Nicaragua. The World Bank Economic Review, 17(2), 283–295. Sabia, J. J. (2009). School-year employment and academic performance of young adolescents. Economics of Education Review, 28(2), 268–276. Sedlacek, G., Duryea, S., Ilahi, N., & Sasaki, M. (2005). Child labor, schooling, and poverty in Latin America. World Bank Social Protection Discussion Paper Series, 51, 1. United Nations Educational, Scientific and Cultural Organization. (2005). EFA global monitoring report 2006. Education for all: Literacy for life. Paris: UNESCO. United Nations Educational, Scientific and Cultural Organization. (2010). EFA global monitoring report 2010: Reaching the marginalized. UNESCO/Oxford University Press.